GAN Compression

This video explains the paper "GAN Compression: Efficient Architectures for Interactive Conditional GANs"! This technique adapts Knowledge Distillation for the GAN framework by copying intermediate features from the teacher to student generator, transferring the pre-trained teacher discriminator, and structuring image-to-image translation problems in the "paired" setting by using the teacher generator image as the ground truth image. This paper also explores the use of One-Shot Neural Architecture Search to find an efficient architecture for the student generator network! Thanks for watching, Please Subscribe! Paper Links: GAN Compression: https://ift.tt/2QLxjIg GAN Compression Video Demo from Authors: https://www.youtube.com/watch?v=31AhcLqWc68&list=PL80kAHvQbh-r5R8UmXhQK1ndqRvPNw_ex&index=3&t=0s Pix2Pix: https://ift.tt/2kPCvgx CycleGAN: https://ift.tt/2opD3rk GauGAN: https://ift.tt/2CsbLsZ AVID: https://ift.tt/2xoCvLo DermGAN: https://ift.tt/2vwCAMt SimGAN: https://ift.tt/2ohC3bJ MobileNets: https://ift.tt/2o1bEiR One-Shot Neural Architecture Search: https://ift.tt/2JgmmdK Intro Music: "Runs" from Unminus Thanks for watching! Please Subscribe!